Junshi Finance mainly focuses on fields such as logic circuits, software, online advertising, essential/optional consumption, etc., updating valuable information related to fundamentals daily.
Core Insights from ASIC Server Industry ResearchQ: Recently, there has been a lot of news about OpenAI, including collaborations with NVIDIA, AMD, and Arm. Can you update us on the current overall demand situation for OpenAI and the progress of the company’s projects in collaboration with OpenAI?A: OpenAI currently has multiple ecological cooperation lines, all of which have been publicly announced. One is the ecological cooperation with Broadcom for self-developed ASIC chips, where the company is the sole designer and manufacturer. They secured the project segment from late May to early June and are currently in the R&D phase. The project is planned to complete NPI (New Product Introduction) by June 30, 2026, and enter MP (Mass Production), with expected output in the second half of 2026. The annual planned output is about 2000 cabinets, expected to be realized between the first half of 2027 and the first quarter of 2028. The chip density per cabinet is 128 chips, with a single cabinet valued at approximately $3 million, which includes 128 ASIC chips. This is currently the most direct and complete project the company has with OpenAI. The second is a collaboration with AMD to build a 6GW data center, expected to have 5-6 centers. The company is striving to participate in the design and manufacturing of the switching part, similar to the collaboration model with AMD on the Meta project, and the scale of the AMD project is expected to be no less than that of its own ecosystem. The third is the collaboration with NVIDIA, involving an investment of $100 billion for several GW data centers, with at least $30 billion invested annually over the next three years. The scale of the ASIC project with OpenAI is about $7-8 billion, while the AMD project is between $7-8 billion and $9 billion, with NVIDIA mainly involved in leasing fees. The fourth is a collaboration with Arm to form a new ecological chain. In this way, OpenAI has established cooperative alliances with all major chip players in the industry.Q: Which company’s CPU is used in the self-developed ASIC project by OpenAI? Is it Arm or AMD?A: The project uses both Arm and AMD versions of the CPU. The first version uses AMD’s CPU along with Broadcom’s ASIC chips, mainly for training. The chips developed by Arm in the future will mainly be used for inference.Q: What is the current progress of the self-developed ASIC project with Broadcom?A: The project is planned to tape out by April 2026 at the latest, originally scheduled for July 2026. Mass production is expected in July 2026, but it may be delayed until January 2027. Currently, all parties involved are accelerating progress because OpenAI has a large plan, and all players are actively cooperating.Q: The annual expected output is 2000 cabinets; what is the specific timeline?A: The 2000 cabinets are expected to start from April 2027 and be completed by March 2028. In the second half of 2025 and 2026, there will be trial production for mass production from around October to the end of March 2027.Q: The value of a single cabinet server is about $3 million; what is the value of the part the company is responsible for?A: The value of the part the company is responsible for is about $1.5 million.Q: What is the value of a single ASIC?A: The value of a single ASIC is about $12,000.Q: What parts is the company responsible for? Does it include all assembly?A: The company will be responsible for all assembly processes except for the chip part.Q: What is the outlook for the gross margin of this project?A: The gross margin has not been disclosed yet, but it is expected not to be higher than the Google project. The competition for this project is fierce, with companies like Foxconn and Wistron also participating. During the bidding process, companies like Foxconn, Wistron, and Accton participated, and the company was ultimately chosen mainly due to its rich experience and capabilities in ASIC processing and design. However, the price requirements from the other party are relatively aggressive. Overall, the gross margin is expected not to exceed Google’s level.Q: What is the cycle and amount of the OpenAI project?A: The project cycle for OpenAI is actually four years, with a total amount of $10 billion. Strictly speaking, the project starts in 2026, and specific output will not begin until the fourth quarter of 2026. It can be understood that the $10 billion will be consumed over four years from 2026 to 2029.Q: Besides ASIC servers, what is the expected procurement demand for switches? Are switches included in the $10 billion project amount over four years?A: Switches are procured separately; for example, a 1.6T switch is considered ordinary procurement and does not fall under the customized procurement in the AI project, so it is not included in the $10 billion project amount over four years.Q: What is OpenAI’s specific procurement demand for switches? What is the current procurement progress?A: Currently, the procurement of 1.6T switches has not yet started, and it is expected to be launched in December next year (2026). The procurement method will adopt the JDM model, with an initial procurement volume of about 2000 units, mainly for trial use. The current largest procurement volume still comes from Google and Meta.Q: Looking ahead to OpenAI’s AMD project, it was mentioned that the expected scale will not be less than its own ecosystem. What are the expected project amount and cycle? How significant will the company’s participation be?A: The AMD project scale is at least three years, and the amount may reach several billion dollars. The company’s algorithms differ from AMD’s; the company only involves CPUs and Broadcom chips, and ASIC chips are not included. OpenAI directly procures AMD’s MI450 chips, and this part of the revenue is not counted in the company’s total assembly and design. Calculating only based on CPUs and design, it will account for at most one-third of AMD’s total procurement annually. AMD will design the computing part itself, while the switching part will require external cooperation. The company is currently a preferred supplier and is striving to take over the entire EMS manufacturing process. Since AMD itself does not have manufacturing capabilities, the acquired New Media only retained the server team, but the factory was sold, so the final system manufacturing will inevitably be outsourced. The project has not yet been fully confirmed to be undertaken by the company, and it is expected that decisions and design initiation must be completed within a quarter, with computing and switching parts needing to advance simultaneously.Q: What is the outlook for the company’s chances of winning the AMD project contract manufacturing?A: The company is very determined about this project and plans to expand its factory in the U.S. by the end of the year, especially the Richardson factory, to re-equip it to undertake Google’s spare parts library and after-sales service center in North America, while providing local production capacity assurance for major customers in North America. Local manufacturing in the U.S. can effectively respond to supply chain and political risks, improving customer communication efficiency. It is expected that within six months, the U.S. factory will be upgraded to the world’s largest super factory.Q: How is the U.S. factory evaluated in terms of cost and production yield? Do customers have a demand for localized production?A: The U.S. factory can initially serve as a secondary assembly base, with components produced in Thailand and then shipped to the U.S. for final assembly to meet local origin certification requirements. More and more customers hope to have suppliers and factories in North America for first-line support, spare parts management, and easy visits. As the importance of the project increases and communication needs grow, local factories are becoming a trend, especially in the AI industry, where policy support and subsidies are also promoting localized production.Q: What is the delivery pace for Meta’s Minerva this year (2025)?A: This year (2025), the original plan was to trial produce 2000 cabinets, but some quality control issues have arisen during implementation, which have persisted for about a month. Meta requires relevant suppliers to provide solutions within a month. If the problem can be resolved, Meta will continue to authorize them for contract manufacturing; if not, the order may revert to the company. Originally, the company planned to complete the design and directly outsource the computing part, only doing the switching part as contract manufacturing. If the problem cannot be resolved, the order may revert to the company to continue doing everything.Q: Are there any adjustments to the expected shipments for Minerva next year (2026)?A: Currently, there are no adjustments, but there may be a final adjustment opportunity before November 15. It mainly depends on the feedback from Quanta. If Quanta’s feedback is positive, Meta will formally consider how many sets to ship next year (2026). Currently, the total for this year (2025) and next year (2026) is 18,000 sets, and in principle, next year (2026) is 16,000 sets. Whether to increase to 20,000 sets or more will wait for Meta’s final evaluation.Q: What is the overall progress of MTIA2?A: It is expected to be in the first quarter of 2027.Q: Has it been delayed?A: Yes, it has been delayed. The reason for the delay is that Meta has now adopted a multi-ecosystem strategy, developing Minerva in-house while collaborating with AMD on Anacapa, and continuing to procure NVIDIA GPUs (GB200, GB300) and Google’s TPUs. With these four lines advancing simultaneously, there will basically be no shortage of computing power next year (2026), so more resources can be invested in optimization on the self-developed line. Originally, MTIA2 was planned to be launched in the third quarter of next year (2026), but after discussions with Broadcom, it was decided to let Broadcom continue optimization, delaying it to the first quarter of 2027, possibly January 2027.Q: What is the scale of the projects between Meta and AMD? How many AMD chips are expected to be procured next year (2026)?A: Currently, there are no specific data; this project started later than Minerva, and the EVT (Engineering Validation Test) has not yet been released. It is known that the switching part uses 8 boards, each with two TH6 chips, and the computing board uses about 18 to 20 boards, with a maximum of 4 chips per board, model MI355. The quantity will be refined based on the results after the EVT engineering prototype is released.Q: When is the EVT expected to be released?A: The first phase should be released between November 10 and the end of November. The company collaborates with AMD, where the company is responsible for the switching part, and AMD is responsible for the computing part, with a collaboration model similar to that of OpenAI and AMD.Q: This year, how will OCP Meta integrate the Spectrum X Ethernet switch into the FBoss network infrastructure, and what impact will it have on the company?A: It mainly serves as a unified design for out-of-cabinet infrastructure, used for basic switching construction, and will not affect the current ASIC part. Meta’s network still includes OEM and ODM, although the conversion rate for ODM is relatively high, about 50%, but half is still OEM. For supply, matching is necessary to achieve transmission. Meta has announced a switch project in collaboration with the company on OCP, with 12 1.6T switches per cabinet, generating about $500-600 million annually, and in the next three years, from 2026 to 2028, there will also be an output of $500-600 million annually.Q: Who is responsible for the PCB of OpenAI’s ASIC servers? Is it TTM?A: OpenAI’s main PCB supplier is currently TTM.Q: How should the company’s value in the Anacapa project be estimated?A: The value of the Anacapa project can be directly estimated based on the value of 8 1.6T switching boards, as the computing part is not the company’s responsibility.
Note:The above content is sourced from the internet and is for personal investment learning record purposes only.
The views expressed by experts in this report do not represent the views or opinions of Junshi Finance. Junshi Finance does not bear any responsibility for the accuracy, reliability, or completeness of the information, nor for any direct or indirect losses or damages arising from the use of this information or any part thereof.
If there is any inconvenience, please contact the backend for deletion; contact information can be found in the public account menu.